Scalable Spatiotemporal Estimation of Particulate Matter via Multi-Source Data Fusion and Deep Learning: A Comprehensive Review

Estimating and forecasting fine particulate matter (PM2.5 and PM10) accurately is critical for public health and protecting the environment. Ground-based measurement networks have high accuracy and precision, but for the rapid urbanisation scenario in India, the coverage is too sparse. To address this problem, the research community has more and more started to combine satellite-based Aerosol Optical Depth (AOD) data and meteorological reanalysis data using Machine Learning (ML) methods. This paper is the first to review the evolution of this technique, outlining the fusion of various data sources. It analyses the use of the low-earth orbiting MODIS and VIIRS sensors, and the more recently used high- frequency geostationary meteorological satellites such as INSAT- 3DR. The review also analyses the atmospheric physics involved in estimating surface PM from column-integrated AOD. The paper also outlines the predictive modelling techniques using traditional statistical regression and more advanced ensemble models such as Random Forest, XGBoost and Long Short-Term Memory networks (LSTM), which are used for spatiotemporal deep learning. Lastly, this paper analyses the extensive research needed in vertical aerosol profiling, spatial nonstationarity and computational time.